Title:
【発明の名称】メンバシップ関数を用いたニューラルネットワーク及びその学習方式
Document Type and Number:
Japanese Patent JP3137996
Kind Code:
B2
Abstract:
PURPOSE:To simplify the learning of a neural network and also to facilitate the connection weight analysis by adding a membership function attaining layer between an input layer and a medium layer and setting previously the weight of the network connection and the neuron threshold value at the part of the membership function attaining layer. CONSTITUTION:A neural network includes an input layer 10 serving as a branch unit, a medium layer 12 of a double layer structure using a neuron, and an output layer 14. Then a membership function attaining layer 16 is added between both layers 10 and 12. The layer 16 consists of a 1st layer of a neuron having a sigmoid function as a nonlinear function and a 2nd layer having a neuron having a linear function. The layer 12 consists of the 3rd and 4th layers, and each neuron of the layer 16 approximates the small, medium and large membership functions.
Inventors:
Yuri Owada
Nobuo Watanabe
Asahi Kawamura
Ryusuke Masuoka
Kazuo Asakawa
Shigenori Matsuoka
Hiroyuki Okada
Nobuo Watanabe
Asahi Kawamura
Ryusuke Masuoka
Kazuo Asakawa
Shigenori Matsuoka
Hiroyuki Okada
Application Number:
JP3990391A
Publication Date:
February 26, 2001
Filing Date:
March 06, 1991
Export Citation:
Assignee:
富士通株式会社
株式会社エフ・エフ・シー
株式会社エフ・エフ・シー
International Classes:
G06F15/18; G06F9/44; G06G7/60; G06N3/00; G06N99/00; (IPC1-7): G06G7/60; G06F15/18; G06N3/00
Domestic Patent References:
| JP464132A | ||||
| JP4275629A | ||||
| JP4275634A |
Attorney, Agent or Firm:
Susumu Takeuchi (1 person outside)